Harnessing Structural Dynamics to Guide Deep Learning: A Physics-Aware Framework for Protein Design

Deep learning models like BayesDesign offer powerful tools for protein sequence optimization, yet their "black-box" nature can yield mutations that compromise enzymatic activity. We present a physics-aware framework that guides deep learning predictions by incorporating structural and dynamic constraints, demonstrated through the engineering of the Nanoluc luciferase. Rather than allowing unrestricted mutations, we constrained BayesDesign by selecting specific sites based on biophysical principles. This included targeting mutations to disordered loop regions distal to the catalytic center and using molecular dynamics simulations with shortest-path analysis to identify and preserve critical residue interactions.



This expert-guided approach successfully navigated the complex stability-activity landscape of the chosen luciferase. We generated variants preserving native flexibility, evidenced by an inverse relationship between per-residue AlphaFold3 confidence scores (pLDDT) and molecular dynamics-derived atomic fluctuations (RMSF), indicating alignment between predicted structural uncertainty and observed flexibility. Experimental validation of these variants demonstrated enhanced thermostability and a catalytic activity over 3-fold higher than wild-type at elevated temperatures. This work establishes a robust, generalizable strategy for protein engineering that integrates the predictive power of deep learning with fundamental principles of protein physics.